2 * Copyright(c) 2015 - 2017 Intel Corporation.
4 * This file is provided under a dual BSD/GPLv2 license. When using or
5 * redistributing this file, you may do so under either license.
9 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of version 2 of the GNU General Public License as
11 * published by the Free Software Foundation.
13 * This program is distributed in the hope that it will be useful, but
14 * WITHOUT ANY WARRANTY; without even the implied warranty of
15 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
16 * General Public License for more details.
20 * Redistribution and use in source and binary forms, with or without
21 * modification, are permitted provided that the following conditions
24 * - Redistributions of source code must retain the above copyright
25 * notice, this list of conditions and the following disclaimer.
26 * - Redistributions in binary form must reproduce the above copyright
27 * notice, this list of conditions and the following disclaimer in
28 * the documentation and/or other materials provided with the
30 * - Neither the name of Intel Corporation nor the names of its
31 * contributors may be used to endorse or promote products derived
32 * from this software without specific prior written permission.
34 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
35 * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
36 * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR
37 * A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT
38 * OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT, INCIDENTAL,
39 * SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING, BUT NOT
40 * LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS OF USE,
41 * DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND ON ANY
42 * THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
43 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
44 * OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
48 #include <linux/types.h>
49 #include <linux/device.h>
50 #include <linux/dmapool.h>
51 #include <linux/slab.h>
52 #include <linux/list.h>
53 #include <linux/highmem.h>
55 #include <linux/uio.h>
56 #include <linux/rbtree.h>
57 #include <linux/spinlock.h>
58 #include <linux/delay.h>
59 #include <linux/kthread.h>
60 #include <linux/mmu_context.h>
61 #include <linux/module.h>
62 #include <linux/vmalloc.h>
63 #include <linux/string.h>
67 #include "user_sdma.h"
68 #include "verbs.h" /* for the headers */
69 #include "common.h" /* for struct hfi1_tid_info */
73 static uint hfi1_sdma_comp_ring_size = 128;
74 module_param_named(sdma_comp_size, hfi1_sdma_comp_ring_size, uint, S_IRUGO);
75 MODULE_PARM_DESC(sdma_comp_size, "Size of User SDMA completion ring. Default: 128");
77 /* The maximum number of Data io vectors per message/request */
78 #define MAX_VECTORS_PER_REQ 8
80 * Maximum number of packet to send from each message/request
81 * before moving to the next one.
83 #define MAX_PKTS_PER_QUEUE 16
85 #define num_pages(x) (1 + ((((x) - 1) & PAGE_MASK) >> PAGE_SHIFT))
87 #define req_opcode(x) \
88 (((x) >> HFI1_SDMA_REQ_OPCODE_SHIFT) & HFI1_SDMA_REQ_OPCODE_MASK)
89 #define req_version(x) \
90 (((x) >> HFI1_SDMA_REQ_VERSION_SHIFT) & HFI1_SDMA_REQ_OPCODE_MASK)
91 #define req_iovcnt(x) \
92 (((x) >> HFI1_SDMA_REQ_IOVCNT_SHIFT) & HFI1_SDMA_REQ_IOVCNT_MASK)
94 /* Number of BTH.PSN bits used for sequence number in expected rcvs */
95 #define BTH_SEQ_MASK 0x7ffull
97 #define AHG_KDETH_INTR_SHIFT 12
98 #define AHG_KDETH_SH_SHIFT 13
99 #define AHG_KDETH_ARRAY_SIZE 9
101 #define PBC2LRH(x) ((((x) & 0xfff) << 2) - 4)
102 #define LRH2PBC(x) ((((x) >> 2) + 1) & 0xfff)
104 #define AHG_HEADER_SET(arr, idx, dw, bit, width, value) \
106 if ((idx) < ARRAY_SIZE((arr))) \
107 (arr)[(idx++)] = sdma_build_ahg_descriptor( \
108 (__force u16)(value), (dw), (bit), \
114 /* Tx request flag bits */
115 #define TXREQ_FLAGS_REQ_ACK BIT(0) /* Set the ACK bit in the header */
116 #define TXREQ_FLAGS_REQ_DISABLE_SH BIT(1) /* Disable header suppression */
118 #define SDMA_PKT_Q_INACTIVE BIT(0)
119 #define SDMA_PKT_Q_ACTIVE BIT(1)
120 #define SDMA_PKT_Q_DEFERRED BIT(2)
123 * Maximum retry attempts to submit a TX request
124 * before putting the process to sleep.
126 #define MAX_DEFER_RETRY_COUNT 1
128 static unsigned initial_pkt_count = 8;
130 #define SDMA_IOWAIT_TIMEOUT 1000 /* in milliseconds */
132 struct sdma_mmu_node;
134 struct user_sdma_iovec {
135 struct list_head list;
137 /* number of pages in this vector */
139 /* array of pinned pages for this vector */
142 * offset into the virtual address space of the vector at
143 * which we last left off.
146 struct sdma_mmu_node *node;
149 struct sdma_mmu_node {
150 struct mmu_rb_node rb;
151 struct hfi1_user_sdma_pkt_q *pq;
157 /* evict operation argument */
159 u32 cleared; /* count evicted so far */
160 u32 target; /* target count to evict */
163 struct user_sdma_request {
164 /* This is the original header from user space */
165 struct hfi1_pkt_header hdr;
167 /* Read mostly fields */
168 struct hfi1_user_sdma_pkt_q *pq ____cacheline_aligned_in_smp;
169 struct hfi1_user_sdma_comp_q *cq;
171 * Pointer to the SDMA engine for this request.
172 * Since different request could be on different VLs,
173 * each request will need it's own engine pointer.
175 struct sdma_engine *sde;
176 struct sdma_req_info info;
177 /* TID array values copied from the tid_iov vector */
179 /* total length of the data in the request */
181 /* number of elements copied to the tids array */
184 * We copy the iovs for this request (based on
185 * info.iovcnt). These are only the data vectors
190 /* Writeable fields shared with interrupt */
191 u64 seqcomp ____cacheline_aligned_in_smp;
193 /* status of the last txreq completed */
196 /* Send side fields */
197 struct list_head txps ____cacheline_aligned_in_smp;
200 * KDETH.OFFSET (TID) field
201 * The offset can cover multiple packets, depending on the
202 * size of the TID entry.
206 * KDETH.Offset (Eager) field
207 * We need to remember the initial value so the headers
208 * can be updated properly.
212 /* TID index copied from the tid_iov vector */
214 /* progress index moving along the iovs array */
219 struct user_sdma_iovec iovs[MAX_VECTORS_PER_REQ];
220 } ____cacheline_aligned_in_smp;
223 * A single txreq could span up to 3 physical pages when the MTU
224 * is sufficiently large (> 4K). Each of the IOV pointers also
225 * needs it's own set of flags so the vector has been handled
226 * independently of each other.
228 struct user_sdma_txreq {
229 /* Packet header for the txreq */
230 struct hfi1_pkt_header hdr;
231 struct sdma_txreq txreq;
232 struct list_head list;
233 struct user_sdma_request *req;
239 #define SDMA_DBG(req, fmt, ...) \
240 hfi1_cdbg(SDMA, "[%u:%u:%u:%u] " fmt, (req)->pq->dd->unit, \
241 (req)->pq->ctxt, (req)->pq->subctxt, (req)->info.comp_idx, \
243 #define SDMA_Q_DBG(pq, fmt, ...) \
244 hfi1_cdbg(SDMA, "[%u:%u:%u] " fmt, (pq)->dd->unit, (pq)->ctxt, \
245 (pq)->subctxt, ##__VA_ARGS__)
247 static int user_sdma_send_pkts(struct user_sdma_request *req,
249 static int num_user_pages(const struct iovec *iov);
250 static void user_sdma_txreq_cb(struct sdma_txreq *txreq, int status);
251 static inline void pq_update(struct hfi1_user_sdma_pkt_q *pq);
252 static void user_sdma_free_request(struct user_sdma_request *req, bool unpin);
253 static int pin_vector_pages(struct user_sdma_request *req,
254 struct user_sdma_iovec *iovec);
255 static void unpin_vector_pages(struct mm_struct *mm, struct page **pages,
256 unsigned start, unsigned npages);
257 static int check_header_template(struct user_sdma_request *req,
258 struct hfi1_pkt_header *hdr, u32 lrhlen,
260 static int set_txreq_header(struct user_sdma_request *req,
261 struct user_sdma_txreq *tx, u32 datalen);
262 static int set_txreq_header_ahg(struct user_sdma_request *req,
263 struct user_sdma_txreq *tx, u32 len);
264 static inline void set_comp_state(struct hfi1_user_sdma_pkt_q *pq,
265 struct hfi1_user_sdma_comp_q *cq,
266 u16 idx, enum hfi1_sdma_comp_state state,
268 static inline u32 set_pkt_bth_psn(__be32 bthpsn, u8 expct, u32 frags);
269 static inline u32 get_lrh_len(struct hfi1_pkt_header, u32 len);
271 static int defer_packet_queue(
272 struct sdma_engine *sde,
274 struct sdma_txreq *txreq,
276 static void activate_packet_queue(struct iowait *wait, int reason);
277 static bool sdma_rb_filter(struct mmu_rb_node *node, unsigned long addr,
279 static int sdma_rb_insert(void *arg, struct mmu_rb_node *mnode);
280 static int sdma_rb_evict(void *arg, struct mmu_rb_node *mnode,
281 void *arg2, bool *stop);
282 static void sdma_rb_remove(void *arg, struct mmu_rb_node *mnode);
283 static int sdma_rb_invalidate(void *arg, struct mmu_rb_node *mnode);
285 static struct mmu_rb_ops sdma_rb_ops = {
286 .filter = sdma_rb_filter,
287 .insert = sdma_rb_insert,
288 .evict = sdma_rb_evict,
289 .remove = sdma_rb_remove,
290 .invalidate = sdma_rb_invalidate
293 static int defer_packet_queue(
294 struct sdma_engine *sde,
296 struct sdma_txreq *txreq,
299 struct hfi1_user_sdma_pkt_q *pq =
300 container_of(wait, struct hfi1_user_sdma_pkt_q, busy);
301 struct hfi1_ibdev *dev = &pq->dd->verbs_dev;
302 struct user_sdma_txreq *tx =
303 container_of(txreq, struct user_sdma_txreq, txreq);
305 if (sdma_progress(sde, seq, txreq)) {
306 if (tx->busycount++ < MAX_DEFER_RETRY_COUNT)
310 * We are assuming that if the list is enqueued somewhere, it
311 * is to the dmawait list since that is the only place where
312 * it is supposed to be enqueued.
314 xchg(&pq->state, SDMA_PKT_Q_DEFERRED);
315 write_seqlock(&dev->iowait_lock);
316 if (list_empty(&pq->busy.list))
317 list_add_tail(&pq->busy.list, &sde->dmawait);
318 write_sequnlock(&dev->iowait_lock);
324 static void activate_packet_queue(struct iowait *wait, int reason)
326 struct hfi1_user_sdma_pkt_q *pq =
327 container_of(wait, struct hfi1_user_sdma_pkt_q, busy);
328 xchg(&pq->state, SDMA_PKT_Q_ACTIVE);
329 wake_up(&wait->wait_dma);
332 static void sdma_kmem_cache_ctor(void *obj)
334 struct user_sdma_txreq *tx = obj;
336 memset(tx, 0, sizeof(*tx));
339 int hfi1_user_sdma_alloc_queues(struct hfi1_ctxtdata *uctxt,
340 struct hfi1_filedata *fd)
344 struct hfi1_devdata *dd;
345 struct hfi1_user_sdma_comp_q *cq;
346 struct hfi1_user_sdma_pkt_q *pq;
352 if (!hfi1_sdma_comp_ring_size)
357 pq = kzalloc(sizeof(*pq), GFP_KERNEL);
361 INIT_LIST_HEAD(&pq->list);
363 pq->ctxt = uctxt->ctxt;
364 pq->subctxt = fd->subctxt;
365 pq->n_max_reqs = hfi1_sdma_comp_ring_size;
366 pq->state = SDMA_PKT_Q_INACTIVE;
367 atomic_set(&pq->n_reqs, 0);
368 init_waitqueue_head(&pq->wait);
369 atomic_set(&pq->n_locked, 0);
372 iowait_init(&pq->busy, 0, NULL, defer_packet_queue,
373 activate_packet_queue, NULL);
376 pq->reqs = kcalloc(hfi1_sdma_comp_ring_size,
382 pq->req_in_use = kcalloc(BITS_TO_LONGS(hfi1_sdma_comp_ring_size),
383 sizeof(*pq->req_in_use),
386 goto pq_reqs_no_in_use;
388 snprintf(buf, 64, "txreq-kmem-cache-%u-%u-%u", dd->unit, uctxt->ctxt,
390 pq->txreq_cache = kmem_cache_create(buf,
391 sizeof(struct user_sdma_txreq),
394 sdma_kmem_cache_ctor);
395 if (!pq->txreq_cache) {
396 dd_dev_err(dd, "[%u] Failed to allocate TxReq cache\n",
401 cq = kzalloc(sizeof(*cq), GFP_KERNEL);
405 cq->comps = vmalloc_user(PAGE_ALIGN(sizeof(*cq->comps)
406 * hfi1_sdma_comp_ring_size));
410 cq->nentries = hfi1_sdma_comp_ring_size;
412 ret = hfi1_mmu_rb_register(pq, pq->mm, &sdma_rb_ops, dd->pport->hfi1_wq,
415 dd_dev_err(dd, "Failed to register with MMU %d", ret);
422 spin_lock_irqsave(&uctxt->sdma_qlock, flags);
423 list_add(&pq->list, &uctxt->sdma_queues);
424 spin_unlock_irqrestore(&uctxt->sdma_qlock, flags);
433 kmem_cache_destroy(pq->txreq_cache);
435 kfree(pq->req_in_use);
444 int hfi1_user_sdma_free_queues(struct hfi1_filedata *fd)
446 struct hfi1_ctxtdata *uctxt = fd->uctxt;
447 struct hfi1_user_sdma_pkt_q *pq;
450 hfi1_cdbg(SDMA, "[%u:%u:%u] Freeing user SDMA queues", uctxt->dd->unit,
451 uctxt->ctxt, fd->subctxt);
455 hfi1_mmu_rb_unregister(pq->handler);
456 spin_lock_irqsave(&uctxt->sdma_qlock, flags);
457 if (!list_empty(&pq->list))
458 list_del_init(&pq->list);
459 spin_unlock_irqrestore(&uctxt->sdma_qlock, flags);
460 iowait_sdma_drain(&pq->busy);
461 /* Wait until all requests have been freed. */
462 wait_event_interruptible(
464 (ACCESS_ONCE(pq->state) == SDMA_PKT_Q_INACTIVE));
466 kfree(pq->req_in_use);
467 kmem_cache_destroy(pq->txreq_cache);
472 vfree(fd->cq->comps);
479 static u8 dlid_to_selector(u16 dlid)
481 static u8 mapping[256];
482 static int initialized;
487 memset(mapping, 0xFF, 256);
491 hash = ((dlid >> 8) ^ dlid) & 0xFF;
492 if (mapping[hash] == 0xFF) {
493 mapping[hash] = next;
494 next = (next + 1) & 0x7F;
497 return mapping[hash];
500 int hfi1_user_sdma_process_request(struct hfi1_filedata *fd,
501 struct iovec *iovec, unsigned long dim,
502 unsigned long *count)
505 struct hfi1_ctxtdata *uctxt = fd->uctxt;
506 struct hfi1_user_sdma_pkt_q *pq = fd->pq;
507 struct hfi1_user_sdma_comp_q *cq = fd->cq;
508 struct hfi1_devdata *dd = pq->dd;
509 unsigned long idx = 0;
510 u8 pcount = initial_pkt_count;
511 struct sdma_req_info info;
512 struct user_sdma_request *req;
518 if (iovec[idx].iov_len < sizeof(info) + sizeof(req->hdr)) {
521 "[%u:%u:%u] First vector not big enough for header %lu/%lu",
522 dd->unit, uctxt->ctxt, fd->subctxt,
523 iovec[idx].iov_len, sizeof(info) + sizeof(req->hdr));
526 ret = copy_from_user(&info, iovec[idx].iov_base, sizeof(info));
528 hfi1_cdbg(SDMA, "[%u:%u:%u] Failed to copy info QW (%d)",
529 dd->unit, uctxt->ctxt, fd->subctxt, ret);
533 trace_hfi1_sdma_user_reqinfo(dd, uctxt->ctxt, fd->subctxt,
536 if (info.comp_idx >= hfi1_sdma_comp_ring_size) {
538 "[%u:%u:%u:%u] Invalid comp index",
539 dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx);
544 * Sanity check the header io vector count. Need at least 1 vector
545 * (header) and cannot be larger than the actual io vector count.
547 if (req_iovcnt(info.ctrl) < 1 || req_iovcnt(info.ctrl) > dim) {
549 "[%u:%u:%u:%u] Invalid iov count %d, dim %ld",
550 dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx,
551 req_iovcnt(info.ctrl), dim);
555 if (!info.fragsize) {
557 "[%u:%u:%u:%u] Request does not specify fragsize",
558 dd->unit, uctxt->ctxt, fd->subctxt, info.comp_idx);
562 /* Try to claim the request. */
563 if (test_and_set_bit(info.comp_idx, pq->req_in_use)) {
564 hfi1_cdbg(SDMA, "[%u:%u:%u] Entry %u is in use",
565 dd->unit, uctxt->ctxt, fd->subctxt,
570 * All safety checks have been done and this request has been claimed.
572 hfi1_cdbg(SDMA, "[%u:%u:%u] Using req/comp entry %u\n", dd->unit,
573 uctxt->ctxt, fd->subctxt, info.comp_idx);
574 req = pq->reqs + info.comp_idx;
575 req->data_iovs = req_iovcnt(info.ctrl) - 1; /* subtract header vector */
585 req->seqsubmitted = 0;
589 INIT_LIST_HEAD(&req->txps);
591 memcpy(&req->info, &info, sizeof(info));
593 if (req_opcode(info.ctrl) == EXPECTED) {
594 /* expected must have a TID info and at least one data vector */
595 if (req->data_iovs < 2) {
597 "Not enough vectors for expected request");
604 if (!info.npkts || req->data_iovs > MAX_VECTORS_PER_REQ) {
605 SDMA_DBG(req, "Too many vectors (%u/%u)", req->data_iovs,
606 MAX_VECTORS_PER_REQ);
610 /* Copy the header from the user buffer */
611 ret = copy_from_user(&req->hdr, iovec[idx].iov_base + sizeof(info),
614 SDMA_DBG(req, "Failed to copy header template (%d)", ret);
619 /* If Static rate control is not enabled, sanitize the header. */
620 if (!HFI1_CAP_IS_USET(STATIC_RATE_CTRL))
623 /* Validate the opcode. Do not trust packets from user space blindly. */
624 opcode = (be32_to_cpu(req->hdr.bth[0]) >> 24) & 0xff;
625 if ((opcode & USER_OPCODE_CHECK_MASK) !=
626 USER_OPCODE_CHECK_VAL) {
627 SDMA_DBG(req, "Invalid opcode (%d)", opcode);
632 * Validate the vl. Do not trust packets from user space blindly.
633 * VL comes from PBC, SC comes from LRH, and the VL needs to
634 * match the SC look up.
636 vl = (le16_to_cpu(req->hdr.pbc[0]) >> 12) & 0xF;
637 sc = (((be16_to_cpu(req->hdr.lrh[0]) >> 12) & 0xF) |
638 (((le16_to_cpu(req->hdr.pbc[1]) >> 14) & 0x1) << 4));
639 if (vl >= dd->pport->vls_operational ||
640 vl != sc_to_vlt(dd, sc)) {
641 SDMA_DBG(req, "Invalid SC(%u)/VL(%u)", sc, vl);
646 /* Checking P_KEY for requests from user-space */
647 if (egress_pkey_check(dd->pport, req->hdr.lrh, req->hdr.bth, sc,
648 PKEY_CHECK_INVALID)) {
654 * Also should check the BTH.lnh. If it says the next header is GRH then
655 * the RXE parsing will be off and will land in the middle of the KDETH
656 * or miss it entirely.
658 if ((be16_to_cpu(req->hdr.lrh[0]) & 0x3) == HFI1_LRH_GRH) {
659 SDMA_DBG(req, "User tried to pass in a GRH");
664 req->koffset = le32_to_cpu(req->hdr.kdeth.swdata[6]);
666 * Calculate the initial TID offset based on the values of
667 * KDETH.OFFSET and KDETH.OM that are passed in.
669 req->tidoffset = KDETH_GET(req->hdr.kdeth.ver_tid_offset, OFFSET) *
670 (KDETH_GET(req->hdr.kdeth.ver_tid_offset, OM) ?
671 KDETH_OM_LARGE : KDETH_OM_SMALL);
672 SDMA_DBG(req, "Initial TID offset %u", req->tidoffset);
675 /* Save all the IO vector structures */
676 for (i = 0; i < req->data_iovs; i++) {
677 req->iovs[i].offset = 0;
678 INIT_LIST_HEAD(&req->iovs[i].list);
679 memcpy(&req->iovs[i].iov,
681 sizeof(req->iovs[i].iov));
682 ret = pin_vector_pages(req, &req->iovs[i]);
688 req->data_len += req->iovs[i].iov.iov_len;
690 SDMA_DBG(req, "total data length %u", req->data_len);
692 if (pcount > req->info.npkts)
693 pcount = req->info.npkts;
696 * User space will provide the TID info only when the
697 * request type is EXPECTED. This is true even if there is
698 * only one packet in the request and the header is already
699 * setup. The reason for the singular TID case is that the
700 * driver needs to perform safety checks.
702 if (req_opcode(req->info.ctrl) == EXPECTED) {
703 u16 ntids = iovec[idx].iov_len / sizeof(*req->tids);
706 if (!ntids || ntids > MAX_TID_PAIR_ENTRIES) {
712 * We have to copy all of the tids because they may vary
713 * in size and, therefore, the TID count might not be
714 * equal to the pkt count. However, there is no way to
715 * tell at this point.
717 tmp = memdup_user(iovec[idx].iov_base,
718 ntids * sizeof(*req->tids));
721 SDMA_DBG(req, "Failed to copy %d TIDs (%d)",
731 dlid = be16_to_cpu(req->hdr.lrh[1]);
732 selector = dlid_to_selector(dlid);
733 selector += uctxt->ctxt + fd->subctxt;
734 req->sde = sdma_select_user_engine(dd, selector, vl);
736 if (!req->sde || !sdma_running(req->sde)) {
741 /* We don't need an AHG entry if the request contains only one packet */
742 if (req->info.npkts > 1 && HFI1_CAP_IS_USET(SDMA_AHG))
743 req->ahg_idx = sdma_ahg_alloc(req->sde);
745 set_comp_state(pq, cq, info.comp_idx, QUEUED, 0);
746 atomic_inc(&pq->n_reqs);
748 /* Send the first N packets in the request to buy us some time */
749 ret = user_sdma_send_pkts(req, pcount);
750 if (unlikely(ret < 0 && ret != -EBUSY)) {
756 * It is possible that the SDMA engine would have processed all the
757 * submitted packets by the time we get here. Therefore, only set
758 * packet queue state to ACTIVE if there are still uncompleted
761 if (atomic_read(&pq->n_reqs))
762 xchg(&pq->state, SDMA_PKT_Q_ACTIVE);
765 * This is a somewhat blocking send implementation.
766 * The driver will block the caller until all packets of the
767 * request have been submitted to the SDMA engine. However, it
768 * will not wait for send completions.
770 while (req->seqsubmitted != req->info.npkts) {
771 ret = user_sdma_send_pkts(req, pcount);
775 WRITE_ONCE(req->has_error, 1);
776 if (ACCESS_ONCE(req->seqcomp) ==
777 req->seqsubmitted - 1)
781 wait_event_interruptible_timeout(
783 (pq->state == SDMA_PKT_Q_ACTIVE),
785 SDMA_IOWAIT_TIMEOUT));
791 user_sdma_free_request(req, true);
794 set_comp_state(pq, cq, info.comp_idx, ERROR, req->status);
798 static inline u32 compute_data_length(struct user_sdma_request *req,
799 struct user_sdma_txreq *tx)
802 * Determine the proper size of the packet data.
803 * The size of the data of the first packet is in the header
804 * template. However, it includes the header and ICRC, which need
806 * The minimum representable packet data length in a header is 4 bytes,
807 * therefore, when the data length request is less than 4 bytes, there's
808 * only one packet, and the packet data length is equal to that of the
809 * request data length.
810 * The size of the remaining packets is the minimum of the frag
811 * size (MTU) or remaining data in the request.
816 if (req->data_len < sizeof(u32))
819 len = ((be16_to_cpu(req->hdr.lrh[2]) << 2) -
820 (sizeof(tx->hdr) - 4));
821 } else if (req_opcode(req->info.ctrl) == EXPECTED) {
822 u32 tidlen = EXP_TID_GET(req->tids[req->tididx], LEN) *
825 * Get the data length based on the remaining space in the
828 len = min(tidlen - req->tidoffset, (u32)req->info.fragsize);
829 /* If we've filled up the TID pair, move to the next one. */
830 if (unlikely(!len) && ++req->tididx < req->n_tids &&
831 req->tids[req->tididx]) {
832 tidlen = EXP_TID_GET(req->tids[req->tididx],
835 len = min_t(u32, tidlen, req->info.fragsize);
838 * Since the TID pairs map entire pages, make sure that we
839 * are not going to try to send more data that we have
842 len = min(len, req->data_len - req->sent);
844 len = min(req->data_len - req->sent, (u32)req->info.fragsize);
846 SDMA_DBG(req, "Data Length = %u", len);
850 static inline u32 pad_len(u32 len)
852 if (len & (sizeof(u32) - 1))
853 len += sizeof(u32) - (len & (sizeof(u32) - 1));
857 static inline u32 get_lrh_len(struct hfi1_pkt_header hdr, u32 len)
859 /* (Size of complete header - size of PBC) + 4B ICRC + data length */
860 return ((sizeof(hdr) - sizeof(hdr.pbc)) + 4 + len);
863 static int user_sdma_send_pkts(struct user_sdma_request *req, unsigned maxpkts)
867 struct user_sdma_txreq *tx = NULL;
868 struct hfi1_user_sdma_pkt_q *pq = NULL;
869 struct user_sdma_iovec *iovec = NULL;
876 /* If tx completion has reported an error, we are done. */
877 if (READ_ONCE(req->has_error))
881 * Check if we might have sent the entire request already
883 if (unlikely(req->seqnum == req->info.npkts)) {
884 if (!list_empty(&req->txps))
889 if (!maxpkts || maxpkts > req->info.npkts - req->seqnum)
890 maxpkts = req->info.npkts - req->seqnum;
892 while (npkts < maxpkts) {
893 u32 datalen = 0, queued = 0, data_sent = 0;
897 * Check whether any of the completions have come back
898 * with errors. If so, we are not going to process any
899 * more packets from this request.
901 if (READ_ONCE(req->has_error))
904 tx = kmem_cache_alloc(pq->txreq_cache, GFP_KERNEL);
911 INIT_LIST_HEAD(&tx->list);
914 * For the last packet set the ACK request
915 * and disable header suppression.
917 if (req->seqnum == req->info.npkts - 1)
918 tx->flags |= (TXREQ_FLAGS_REQ_ACK |
919 TXREQ_FLAGS_REQ_DISABLE_SH);
922 * Calculate the payload size - this is min of the fragment
923 * (MTU) size or the remaining bytes in the request but only
924 * if we have payload data.
927 iovec = &req->iovs[req->iov_idx];
928 if (ACCESS_ONCE(iovec->offset) == iovec->iov.iov_len) {
929 if (++req->iov_idx == req->data_iovs) {
933 iovec = &req->iovs[req->iov_idx];
934 WARN_ON(iovec->offset);
937 datalen = compute_data_length(req, tx);
940 * Disable header suppression for the payload <= 8DWS.
941 * If there is an uncorrectable error in the receive
942 * data FIFO when the received payload size is less than
943 * or equal to 8DWS then the RxDmaDataFifoRdUncErr is
944 * not reported.There is set RHF.EccErr if the header
949 "Request has data but pkt len is 0");
952 } else if (datalen <= 32) {
953 tx->flags |= TXREQ_FLAGS_REQ_DISABLE_SH;
957 if (req->ahg_idx >= 0) {
959 u16 pbclen = le16_to_cpu(req->hdr.pbc[0]);
960 u32 lrhlen = get_lrh_len(req->hdr,
963 * Copy the request header into the tx header
964 * because the HW needs a cacheline-aligned
966 * This copy can be optimized out if the hdr
967 * member of user_sdma_request were also
970 memcpy(&tx->hdr, &req->hdr, sizeof(tx->hdr));
971 if (PBC2LRH(pbclen) != lrhlen) {
972 pbclen = (pbclen & 0xf000) |
974 tx->hdr.pbc[0] = cpu_to_le16(pbclen);
976 ret = check_header_template(req, &tx->hdr,
980 ret = sdma_txinit_ahg(&tx->txreq,
981 SDMA_TXREQ_F_AHG_COPY,
982 sizeof(tx->hdr) + datalen,
983 req->ahg_idx, 0, NULL, 0,
987 ret = sdma_txadd_kvaddr(pq->dd, &tx->txreq,
995 changes = set_txreq_header_ahg(req, tx,
1001 ret = sdma_txinit(&tx->txreq, 0, sizeof(req->hdr) +
1002 datalen, user_sdma_txreq_cb);
1006 * Modify the header for this packet. This only needs
1007 * to be done if we are not going to use AHG. Otherwise,
1008 * the HW will do it based on the changes we gave it
1009 * during sdma_txinit_ahg().
1011 ret = set_txreq_header(req, tx, datalen);
1017 * If the request contains any data vectors, add up to
1018 * fragsize bytes to the descriptor.
1020 while (queued < datalen &&
1021 (req->sent + data_sent) < req->data_len) {
1022 unsigned long base, offset;
1023 unsigned pageidx, len;
1025 base = (unsigned long)iovec->iov.iov_base;
1026 offset = offset_in_page(base + iovec->offset +
1028 pageidx = (((iovec->offset + iov_offset +
1029 base) - (base & PAGE_MASK)) >> PAGE_SHIFT);
1030 len = offset + req->info.fragsize > PAGE_SIZE ?
1031 PAGE_SIZE - offset : req->info.fragsize;
1032 len = min((datalen - queued), len);
1033 ret = sdma_txadd_page(pq->dd, &tx->txreq,
1034 iovec->pages[pageidx],
1037 SDMA_DBG(req, "SDMA txreq add page failed %d\n",
1044 if (unlikely(queued < datalen &&
1045 pageidx == iovec->npages &&
1046 req->iov_idx < req->data_iovs - 1)) {
1047 iovec->offset += iov_offset;
1048 iovec = &req->iovs[++req->iov_idx];
1053 * The txreq was submitted successfully so we can update
1056 req->koffset += datalen;
1057 if (req_opcode(req->info.ctrl) == EXPECTED)
1058 req->tidoffset += datalen;
1059 req->sent += data_sent;
1061 iovec->offset += iov_offset;
1062 list_add_tail(&tx->txreq.list, &req->txps);
1064 * It is important to increment this here as it is used to
1065 * generate the BTH.PSN and, therefore, can't be bulk-updated
1066 * outside of the loop.
1068 tx->seqnum = req->seqnum++;
1072 ret = sdma_send_txlist(req->sde, &pq->busy, &req->txps, &count);
1073 req->seqsubmitted += count;
1074 if (req->seqsubmitted == req->info.npkts) {
1075 WRITE_ONCE(req->done, 1);
1077 * The txreq has already been submitted to the HW queue
1078 * so we can free the AHG entry now. Corruption will not
1079 * happen due to the sequential manner in which
1080 * descriptors are processed.
1082 if (req->ahg_idx >= 0)
1083 sdma_ahg_free(req->sde, req->ahg_idx);
1088 sdma_txclean(pq->dd, &tx->txreq);
1090 kmem_cache_free(pq->txreq_cache, tx);
1095 * How many pages in this iovec element?
1097 static inline int num_user_pages(const struct iovec *iov)
1099 const unsigned long addr = (unsigned long)iov->iov_base;
1100 const unsigned long len = iov->iov_len;
1101 const unsigned long spage = addr & PAGE_MASK;
1102 const unsigned long epage = (addr + len - 1) & PAGE_MASK;
1104 return 1 + ((epage - spage) >> PAGE_SHIFT);
1107 static u32 sdma_cache_evict(struct hfi1_user_sdma_pkt_q *pq, u32 npages)
1109 struct evict_data evict_data;
1111 evict_data.cleared = 0;
1112 evict_data.target = npages;
1113 hfi1_mmu_rb_evict(pq->handler, &evict_data);
1114 return evict_data.cleared;
1117 static int pin_vector_pages(struct user_sdma_request *req,
1118 struct user_sdma_iovec *iovec)
1120 int ret = 0, pinned, npages, cleared;
1121 struct page **pages;
1122 struct hfi1_user_sdma_pkt_q *pq = req->pq;
1123 struct sdma_mmu_node *node = NULL;
1124 struct mmu_rb_node *rb_node;
1128 hfi1_mmu_rb_remove_unless_exact(pq->handler,
1130 iovec->iov.iov_base,
1131 iovec->iov.iov_len, &rb_node);
1133 node = container_of(rb_node, struct sdma_mmu_node, rb);
1135 atomic_inc(&node->refcount);
1136 iovec->pages = node->pages;
1137 iovec->npages = node->npages;
1144 node = kzalloc(sizeof(*node), GFP_KERNEL);
1148 node->rb.addr = (unsigned long)iovec->iov.iov_base;
1150 atomic_set(&node->refcount, 0);
1153 npages = num_user_pages(&iovec->iov);
1154 if (node->npages < npages) {
1155 pages = kcalloc(npages, sizeof(*pages), GFP_KERNEL);
1157 SDMA_DBG(req, "Failed page array alloc");
1161 memcpy(pages, node->pages, node->npages * sizeof(*pages));
1163 npages -= node->npages;
1166 if (!hfi1_can_pin_pages(pq->dd, pq->mm,
1167 atomic_read(&pq->n_locked), npages)) {
1168 cleared = sdma_cache_evict(pq, npages);
1169 if (cleared >= npages)
1172 pinned = hfi1_acquire_user_pages(pq->mm,
1173 ((unsigned long)iovec->iov.iov_base +
1174 (node->npages * PAGE_SIZE)), npages, 0,
1175 pages + node->npages);
1181 if (pinned != npages) {
1182 unpin_vector_pages(pq->mm, pages, node->npages,
1188 node->rb.len = iovec->iov.iov_len;
1189 node->pages = pages;
1190 node->npages += pinned;
1191 npages = node->npages;
1192 atomic_add(pinned, &pq->n_locked);
1194 iovec->pages = node->pages;
1195 iovec->npages = npages;
1198 ret = hfi1_mmu_rb_insert(req->pq->handler, &node->rb);
1200 atomic_sub(node->npages, &pq->n_locked);
1207 unpin_vector_pages(pq->mm, node->pages, 0, node->npages);
1212 static void unpin_vector_pages(struct mm_struct *mm, struct page **pages,
1213 unsigned start, unsigned npages)
1215 hfi1_release_user_pages(mm, pages + start, npages, false);
1219 static int check_header_template(struct user_sdma_request *req,
1220 struct hfi1_pkt_header *hdr, u32 lrhlen,
1224 * Perform safety checks for any type of packet:
1225 * - transfer size is multiple of 64bytes
1226 * - packet length is multiple of 4 bytes
1227 * - packet length is not larger than MTU size
1229 * These checks are only done for the first packet of the
1230 * transfer since the header is "given" to us by user space.
1231 * For the remainder of the packets we compute the values.
1233 if (req->info.fragsize % PIO_BLOCK_SIZE || lrhlen & 0x3 ||
1234 lrhlen > get_lrh_len(*hdr, req->info.fragsize))
1237 if (req_opcode(req->info.ctrl) == EXPECTED) {
1239 * The header is checked only on the first packet. Furthermore,
1240 * we ensure that at least one TID entry is copied when the
1241 * request is submitted. Therefore, we don't have to verify that
1242 * tididx points to something sane.
1244 u32 tidval = req->tids[req->tididx],
1245 tidlen = EXP_TID_GET(tidval, LEN) * PAGE_SIZE,
1246 tididx = EXP_TID_GET(tidval, IDX),
1247 tidctrl = EXP_TID_GET(tidval, CTRL),
1249 __le32 kval = hdr->kdeth.ver_tid_offset;
1251 tidoff = KDETH_GET(kval, OFFSET) *
1252 (KDETH_GET(req->hdr.kdeth.ver_tid_offset, OM) ?
1253 KDETH_OM_LARGE : KDETH_OM_SMALL);
1255 * Expected receive packets have the following
1256 * additional checks:
1257 * - offset is not larger than the TID size
1258 * - TIDCtrl values match between header and TID array
1259 * - TID indexes match between header and TID array
1261 if ((tidoff + datalen > tidlen) ||
1262 KDETH_GET(kval, TIDCTRL) != tidctrl ||
1263 KDETH_GET(kval, TID) != tididx)
1270 * Correctly set the BTH.PSN field based on type of
1271 * transfer - eager packets can just increment the PSN but
1272 * expected packets encode generation and sequence in the
1273 * BTH.PSN field so just incrementing will result in errors.
1275 static inline u32 set_pkt_bth_psn(__be32 bthpsn, u8 expct, u32 frags)
1277 u32 val = be32_to_cpu(bthpsn),
1278 mask = (HFI1_CAP_IS_KSET(EXTENDED_PSN) ? 0x7fffffffull :
1282 psn = (psn & ~BTH_SEQ_MASK) | ((psn + frags) & BTH_SEQ_MASK);
1288 static int set_txreq_header(struct user_sdma_request *req,
1289 struct user_sdma_txreq *tx, u32 datalen)
1291 struct hfi1_user_sdma_pkt_q *pq = req->pq;
1292 struct hfi1_pkt_header *hdr = &tx->hdr;
1293 u8 omfactor; /* KDETH.OM */
1296 u32 tidval = 0, lrhlen = get_lrh_len(*hdr, pad_len(datalen));
1298 /* Copy the header template to the request before modification */
1299 memcpy(hdr, &req->hdr, sizeof(*hdr));
1302 * Check if the PBC and LRH length are mismatched. If so
1303 * adjust both in the header.
1305 pbclen = le16_to_cpu(hdr->pbc[0]);
1306 if (PBC2LRH(pbclen) != lrhlen) {
1307 pbclen = (pbclen & 0xf000) | LRH2PBC(lrhlen);
1308 hdr->pbc[0] = cpu_to_le16(pbclen);
1309 hdr->lrh[2] = cpu_to_be16(lrhlen >> 2);
1312 * This is the first packet in the sequence that has
1313 * a "static" size that can be used for the rest of
1314 * the packets (besides the last one).
1316 if (unlikely(req->seqnum == 2)) {
1318 * From this point on the lengths in both the
1319 * PBC and LRH are the same until the last
1321 * Adjust the template so we don't have to update
1324 req->hdr.pbc[0] = hdr->pbc[0];
1325 req->hdr.lrh[2] = hdr->lrh[2];
1329 * We only have to modify the header if this is not the
1330 * first packet in the request. Otherwise, we use the
1331 * header given to us.
1333 if (unlikely(!req->seqnum)) {
1334 ret = check_header_template(req, hdr, lrhlen, datalen);
1340 hdr->bth[2] = cpu_to_be32(
1341 set_pkt_bth_psn(hdr->bth[2],
1342 (req_opcode(req->info.ctrl) == EXPECTED),
1345 /* Set ACK request on last packet */
1346 if (unlikely(tx->flags & TXREQ_FLAGS_REQ_ACK))
1347 hdr->bth[2] |= cpu_to_be32(1UL << 31);
1349 /* Set the new offset */
1350 hdr->kdeth.swdata[6] = cpu_to_le32(req->koffset);
1351 /* Expected packets have to fill in the new TID information */
1352 if (req_opcode(req->info.ctrl) == EXPECTED) {
1353 tidval = req->tids[req->tididx];
1355 * If the offset puts us at the end of the current TID,
1356 * advance everything.
1358 if ((req->tidoffset) == (EXP_TID_GET(tidval, LEN) *
1362 * Since we don't copy all the TIDs, all at once,
1363 * we have to check again.
1365 if (++req->tididx > req->n_tids - 1 ||
1366 !req->tids[req->tididx]) {
1369 tidval = req->tids[req->tididx];
1371 omfactor = EXP_TID_GET(tidval, LEN) * PAGE_SIZE >=
1372 KDETH_OM_MAX_SIZE ? KDETH_OM_LARGE_SHIFT :
1373 KDETH_OM_SMALL_SHIFT;
1374 /* Set KDETH.TIDCtrl based on value for this TID. */
1375 KDETH_SET(hdr->kdeth.ver_tid_offset, TIDCTRL,
1376 EXP_TID_GET(tidval, CTRL));
1377 /* Set KDETH.TID based on value for this TID */
1378 KDETH_SET(hdr->kdeth.ver_tid_offset, TID,
1379 EXP_TID_GET(tidval, IDX));
1380 /* Clear KDETH.SH when DISABLE_SH flag is set */
1381 if (unlikely(tx->flags & TXREQ_FLAGS_REQ_DISABLE_SH))
1382 KDETH_SET(hdr->kdeth.ver_tid_offset, SH, 0);
1384 * Set the KDETH.OFFSET and KDETH.OM based on size of
1387 SDMA_DBG(req, "TID offset %ubytes %uunits om%u",
1388 req->tidoffset, req->tidoffset >> omfactor,
1389 omfactor != KDETH_OM_SMALL_SHIFT);
1390 KDETH_SET(hdr->kdeth.ver_tid_offset, OFFSET,
1391 req->tidoffset >> omfactor);
1392 KDETH_SET(hdr->kdeth.ver_tid_offset, OM,
1393 omfactor != KDETH_OM_SMALL_SHIFT);
1396 trace_hfi1_sdma_user_header(pq->dd, pq->ctxt, pq->subctxt,
1397 req->info.comp_idx, hdr, tidval);
1398 return sdma_txadd_kvaddr(pq->dd, &tx->txreq, hdr, sizeof(*hdr));
1401 static int set_txreq_header_ahg(struct user_sdma_request *req,
1402 struct user_sdma_txreq *tx, u32 datalen)
1404 u32 ahg[AHG_KDETH_ARRAY_SIZE];
1406 u8 omfactor; /* KDETH.OM */
1407 struct hfi1_user_sdma_pkt_q *pq = req->pq;
1408 struct hfi1_pkt_header *hdr = &req->hdr;
1409 u16 pbclen = le16_to_cpu(hdr->pbc[0]);
1410 u32 val32, tidval = 0, lrhlen = get_lrh_len(*hdr, pad_len(datalen));
1412 if (PBC2LRH(pbclen) != lrhlen) {
1413 /* PBC.PbcLengthDWs */
1414 AHG_HEADER_SET(ahg, diff, 0, 0, 12,
1415 cpu_to_le16(LRH2PBC(lrhlen)));
1416 /* LRH.PktLen (we need the full 16 bits due to byte swap) */
1417 AHG_HEADER_SET(ahg, diff, 3, 0, 16,
1418 cpu_to_be16(lrhlen >> 2));
1422 * Do the common updates
1424 /* BTH.PSN and BTH.A */
1425 val32 = (be32_to_cpu(hdr->bth[2]) + req->seqnum) &
1426 (HFI1_CAP_IS_KSET(EXTENDED_PSN) ? 0x7fffffff : 0xffffff);
1427 if (unlikely(tx->flags & TXREQ_FLAGS_REQ_ACK))
1429 AHG_HEADER_SET(ahg, diff, 6, 0, 16, cpu_to_be16(val32 >> 16));
1430 AHG_HEADER_SET(ahg, diff, 6, 16, 16, cpu_to_be16(val32 & 0xffff));
1432 AHG_HEADER_SET(ahg, diff, 15, 0, 16,
1433 cpu_to_le16(req->koffset & 0xffff));
1434 AHG_HEADER_SET(ahg, diff, 15, 16, 16, cpu_to_le16(req->koffset >> 16));
1435 if (req_opcode(req->info.ctrl) == EXPECTED) {
1438 tidval = req->tids[req->tididx];
1441 * If the offset puts us at the end of the current TID,
1442 * advance everything.
1444 if ((req->tidoffset) == (EXP_TID_GET(tidval, LEN) *
1448 * Since we don't copy all the TIDs, all at once,
1449 * we have to check again.
1451 if (++req->tididx > req->n_tids - 1 ||
1452 !req->tids[req->tididx])
1454 tidval = req->tids[req->tididx];
1456 omfactor = ((EXP_TID_GET(tidval, LEN) *
1458 KDETH_OM_MAX_SIZE) ? KDETH_OM_LARGE_SHIFT :
1459 KDETH_OM_SMALL_SHIFT;
1460 /* KDETH.OM and KDETH.OFFSET (TID) */
1461 AHG_HEADER_SET(ahg, diff, 7, 0, 16,
1462 ((!!(omfactor - KDETH_OM_SMALL_SHIFT)) << 15 |
1463 ((req->tidoffset >> omfactor)
1465 /* KDETH.TIDCtrl, KDETH.TID, KDETH.Intr, KDETH.SH */
1466 val = cpu_to_le16(((EXP_TID_GET(tidval, CTRL) & 0x3) << 10) |
1467 (EXP_TID_GET(tidval, IDX) & 0x3ff));
1469 if (unlikely(tx->flags & TXREQ_FLAGS_REQ_DISABLE_SH)) {
1470 val |= cpu_to_le16((KDETH_GET(hdr->kdeth.ver_tid_offset,
1472 AHG_KDETH_INTR_SHIFT));
1474 val |= KDETH_GET(hdr->kdeth.ver_tid_offset, SH) ?
1475 cpu_to_le16(0x1 << AHG_KDETH_SH_SHIFT) :
1476 cpu_to_le16((KDETH_GET(hdr->kdeth.ver_tid_offset,
1478 AHG_KDETH_INTR_SHIFT));
1481 AHG_HEADER_SET(ahg, diff, 7, 16, 14, val);
1486 trace_hfi1_sdma_user_header_ahg(pq->dd, pq->ctxt, pq->subctxt,
1487 req->info.comp_idx, req->sde->this_idx,
1488 req->ahg_idx, ahg, diff, tidval);
1489 sdma_txinit_ahg(&tx->txreq,
1490 SDMA_TXREQ_F_USE_AHG,
1491 datalen, req->ahg_idx, diff,
1492 ahg, sizeof(req->hdr),
1493 user_sdma_txreq_cb);
1499 * SDMA tx request completion callback. Called when the SDMA progress
1500 * state machine gets notification that the SDMA descriptors for this
1501 * tx request have been processed by the DMA engine. Called in
1502 * interrupt context.
1504 static void user_sdma_txreq_cb(struct sdma_txreq *txreq, int status)
1506 struct user_sdma_txreq *tx =
1507 container_of(txreq, struct user_sdma_txreq, txreq);
1508 struct user_sdma_request *req;
1509 struct hfi1_user_sdma_pkt_q *pq;
1510 struct hfi1_user_sdma_comp_q *cq;
1520 if (status != SDMA_TXREQ_S_OK) {
1521 SDMA_DBG(req, "SDMA completion with error %d",
1523 WRITE_ONCE(req->has_error, 1);
1526 req->seqcomp = tx->seqnum;
1527 kmem_cache_free(pq->txreq_cache, tx);
1530 idx = req->info.comp_idx;
1531 if (req->status == -1 && status == SDMA_TXREQ_S_OK) {
1532 if (req->seqcomp == req->info.npkts - 1) {
1534 user_sdma_free_request(req, false);
1536 set_comp_state(pq, cq, idx, COMPLETE, 0);
1539 if (status != SDMA_TXREQ_S_OK)
1540 req->status = status;
1541 if (req->seqcomp == (ACCESS_ONCE(req->seqsubmitted) - 1) &&
1542 (READ_ONCE(req->done) ||
1543 READ_ONCE(req->has_error))) {
1544 user_sdma_free_request(req, false);
1546 set_comp_state(pq, cq, idx, ERROR, req->status);
1551 static inline void pq_update(struct hfi1_user_sdma_pkt_q *pq)
1553 if (atomic_dec_and_test(&pq->n_reqs)) {
1554 xchg(&pq->state, SDMA_PKT_Q_INACTIVE);
1559 static void user_sdma_free_request(struct user_sdma_request *req, bool unpin)
1561 if (!list_empty(&req->txps)) {
1562 struct sdma_txreq *t, *p;
1564 list_for_each_entry_safe(t, p, &req->txps, list) {
1565 struct user_sdma_txreq *tx =
1566 container_of(t, struct user_sdma_txreq, txreq);
1567 list_del_init(&t->list);
1568 sdma_txclean(req->pq->dd, t);
1569 kmem_cache_free(req->pq->txreq_cache, tx);
1572 if (req->data_iovs) {
1573 struct sdma_mmu_node *node;
1576 for (i = 0; i < req->data_iovs; i++) {
1577 node = req->iovs[i].node;
1582 hfi1_mmu_rb_remove(req->pq->handler,
1585 atomic_dec(&node->refcount);
1589 clear_bit(req->info.comp_idx, req->pq->req_in_use);
1592 static inline void set_comp_state(struct hfi1_user_sdma_pkt_q *pq,
1593 struct hfi1_user_sdma_comp_q *cq,
1594 u16 idx, enum hfi1_sdma_comp_state state,
1597 hfi1_cdbg(SDMA, "[%u:%u:%u:%u] Setting completion status %u %d",
1598 pq->dd->unit, pq->ctxt, pq->subctxt, idx, state, ret);
1600 cq->comps[idx].errcode = -ret;
1601 smp_wmb(); /* make sure errcode is visible first */
1602 cq->comps[idx].status = state;
1603 trace_hfi1_sdma_user_completion(pq->dd, pq->ctxt, pq->subctxt,
1607 static bool sdma_rb_filter(struct mmu_rb_node *node, unsigned long addr,
1610 return (bool)(node->addr == addr);
1613 static int sdma_rb_insert(void *arg, struct mmu_rb_node *mnode)
1615 struct sdma_mmu_node *node =
1616 container_of(mnode, struct sdma_mmu_node, rb);
1618 atomic_inc(&node->refcount);
1623 * Return 1 to remove the node from the rb tree and call the remove op.
1625 * Called with the rb tree lock held.
1627 static int sdma_rb_evict(void *arg, struct mmu_rb_node *mnode,
1628 void *evict_arg, bool *stop)
1630 struct sdma_mmu_node *node =
1631 container_of(mnode, struct sdma_mmu_node, rb);
1632 struct evict_data *evict_data = evict_arg;
1634 /* is this node still being used? */
1635 if (atomic_read(&node->refcount))
1636 return 0; /* keep this node */
1638 /* this node will be evicted, add its pages to our count */
1639 evict_data->cleared += node->npages;
1641 /* have enough pages been cleared? */
1642 if (evict_data->cleared >= evict_data->target)
1645 return 1; /* remove this node */
1648 static void sdma_rb_remove(void *arg, struct mmu_rb_node *mnode)
1650 struct sdma_mmu_node *node =
1651 container_of(mnode, struct sdma_mmu_node, rb);
1653 atomic_sub(node->npages, &node->pq->n_locked);
1655 unpin_vector_pages(node->pq->mm, node->pages, 0, node->npages);
1660 static int sdma_rb_invalidate(void *arg, struct mmu_rb_node *mnode)
1662 struct sdma_mmu_node *node =
1663 container_of(mnode, struct sdma_mmu_node, rb);
1665 if (!atomic_read(&node->refcount))